Proceedings of the 18th World Congress

The International Federation of Automatic Control

Milano (Italy) August 28 - September 2, 2011

An optimization procedure of the start-up of Combined Cycle Power Plants

F. Casella, M. Farina, F. Righetti, R. Scattolini*
D. Faille, F. Davelaar**
A. Tica , H. Gueguen, D. Dumur***

*Dipartimento di Elettronica e Informazione, Politecnico di Milano,

Via Ponzio 34/5, 20133 Milano, Italy
({casella,farina,scattolini}@elet.polimi.it).
**EDF, quai Watier, Chatou, France ({damien.faille, frans.davelaar} @edf.fr)
*** SUPELEC, France, (e-mail: {adrian.tica, herve.gueguen, didier.dumur}@supelec.fr)

Abstract: This paper presents an optimization procedure for the definition of the gas turbine load profile
during the hot start-up of Combined Cycle Power Plants (CCPP). First a dynamic model of CCPP is
briefly described, together with its implementation in the Modelica language. Then, an identification
procedure is developed to determine a simplified model to be implemented in Matlab/Simulink and to be
used for the solution of the optimization problem. This simplified model is built by interpolating a
number of linear estimated models with local validity. The load profile is assumed to be described by a
suitable function, whose parameters are optimized by solving a minimum time problem subject to the
plant (simulator) dynamics and to a number of constraints to be imposed on the main plant variables,
such as temperatures, pressures, thermal and mechanical stresses. A number of simulation experiments is
reported to witness the performance of the proposed approach.
Keywords: Combined cycle power plants, start-up procedures, control, optimization, simulation, model
identification.

used to produce the input-output data required for the

1. INTRODUCTION
Combined Cycle Power Plants (CCPP) are nowadays widely
diffused for their improved efficiency levels and reduced
pollutant emissions with respect to traditional oil or coal fired
plants, see (Watson, 1996). However, they are often managed
at full load because of significant losses of efficiency at
partial load, even though the increasing energy demand and
grid regulation difficulties due to the energy market
liberalization would require a more flexible management. For
these reasons, it is of interest to develop new and efficient
management strategies, either for control at partial or full
load (Aurora et al., 2005), or for the start-up and shut-down
phases. This second topic is of paramount importance since,
as it is well known, the lifetime of CCPP is mainly related to
the thermal and mechanical stresses reached during the start-
up. For this reason, both in academia and in industry, many
efforts have been devoted to determine procedures reducing
the start-up time while keeping the life-time consumption of
the most critically stressed components under control, see e.g.
(Franke et al, 2003, Krüger et al., 2001, 2004, Bausa and
Tsatsaronis, 2001a, 2001b, Smith et al. 1996, Xu et al. 2000,
Ferrari-Trecate et al., 2002, Casella and Pretolani, 2006,
Albanesi et al. 2006, Faille and Davelaar, 2009).
In this paper, a novel model-based approach to the
optimization of the Gas Turbine (GT) load profile to be used
in the start-up phase is presented. Specifically, first a physical
model of the plant is developed together with a simulation
environment in the Modelica language (Section 2). This
simulator is then considered as the “reference plant” to be
development of a simplified model based on interpolated
locally identified linear models (Section 3). In turn, the
interpolated model is used to compute the optimal load
profile during the start-up by assuming that it can be
described by a parametrised function, whose parameters are
computed by solving a minimum-time optimal control
problem subject to the constraints imposed by the plant
dynamics and by the maximum peak value of the stress
(Section 4). Finally, the optimal load profile is applied to
simulate a start-up procedure on the original detailed
Modelica simulator, so validating the overall project (see
Section 5). Some final remarks and hints for future
developments close the paper (Section 6).
2. PLANT MODEL AND MODELICA SIMULATOR
Many dynamic models of CCPP have been proposed in the
literature and used for control design and for the optimization
of the start-up procedure, see e.g. Aurora, 2004, Aurora et al.
2005 and Casella and Pretolani, 2006. In particular, a detailed
description of the elementary models based on first principle
equations (mass, energy and momentum balances) of the
main elements of a CCPP are reported in Aurora (2004),
where the interested reader is referred to. Models of the
thermal and mechanical stresses have been developed in
Leporati et al. (1993), Lausterer (1997). In Casella and
Pretolani (2006), a detailed simulation model was developed
for a plant composed by a gas turbine unit (GT), coupled to a
heat recovery steam generator (HRSG) with three levels of

978-3-902661-93-7/11/$20.00 © 2011 IFAC 7043 10.3182/20110828-6-IT-1002.00604

18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011
pressure, driving a steam turbine (ST) group. The simulator
was based on the Modelica ThermoPower library, see Casella
and Leva (2003, 2006), and was parametrised with design
and operating data from a typical unit. The simulator was
then validated by replicating a real start-up transient, as
recorded by the plant DCS.
In this work, the simulator described in Casella and Pretolani
(2006) has been adapted to a plant with one level of pressure.
This makes it possible to obtain a more tractable dynamic
model for identification and optimization purposes while
keeping the main characteristics of the problem.
The main components of the simulated plant are:
Gas Turbine: the gas turbine (GT) generates electric power
and hot flue gas. It includes local control and its controlled
input is then its load reference (GTload) expressed as a
percentage of its maximal power. Its characteristics are:
maximal power 235 [MW], nominal flue gas flow rate 585.6
[kg/s] and nominal fuel flow rate 12.1 [kg/s].
Heat Recovery Steam Generator: the heat recovery steam
generator (HRSG) generates steam from the heat of the flue
gas of the gas turbine. Its main components are the
economizer, the steam drum, the evaporator and the
superheater. Its controlled inputs are the drum feedwater flow
rate and the desuperheater water flow rate, this later enabling
to control the temperature of the steam after the superheater.
Its characteristics are: nominal steam flow rate 70.6 [kg/s]
and nominal steam pressure 129.6 [bar].
Steam Turbine: the steam turbine (ST) is a single stage
turbine that generates electric power from the steam. It has
three controlled inputs: the position of the bypass valve, the
position of the admission valve and the generator grid breaker
that connects the turbine to the grid. Its nominal power is 85
[MW].
Condenser: The condenser of the steam that comes from the
ST through the bypass valve or the turbine is modeled, in a
simplified way, as a sink with constant pressure ( 0.1 [bar]).
The variables that are considered of interest by engineers for
the start-up have been defined as outputs of the model and
are:
• power of gas turbine;
• fuel flow in gas turbine;
• drum level of the HRSG;
• temperature of steam in superheater (TSH);
• steam pressure;
• superheater header stress;
• temperature of steam in the steam line;
• power of steam turbine;
• frequency of steam turbine;
• rotor stress of steam turbine;
• steam flow.
This model is designed to study the hot start-up procedure of
the plant. The GTload is then assumed to vary from technical
minimal conditions (7.5% of the full load) to full load
conditions, i.e. from 17.625[MW] to 235[MW]. Local
controllers have been designed to ensure adequate working
conditions for the plant:
7044
• a PI loop that controls the drum feedwater flow in order
to regulate the drum level at a constant value,
• a loop that controls the position of the admission valve
in order to maintain the pressure at a minimum value
(60 bar for instance) that is active when the load is
below 50%.
Finally in this study, only the last stage of the start-up, the
load increase, is considered. It is then assumed that,
• the by-pass valve is closed;
• the steam turbine is connected to the grid;
• the desuperheater feedwater flow gets a value close to
its nominal value (0) and the admission valve is at
minimum opening when the pressure control loop is not
active;
• the load of the gas turbine is set to 15%.
Therefore, the GTload signal is the input that has the biggest
range and induces most of the dynamics characteristics
during the last stage of the start-up.
3. IDENTIFICATION OF INTERPOLATED LOCAL
LINEAR MODELS
Modelica, together with the ThermoPower library (Casella
and Leva 2003, 2006), is an extremely powerful tool for the
rapid development of accurate physical models of thermo
power systems, and in particular of CCPP. On the other hand,
this model is designed for simulation and uses extensively
Modelica features such as genericity and inheritance that
make it unsuitable for optimization. For this reason, it has
been decided to develop a simpler identified model capable to
represent the main dynamics of all the CCPP output variables
of interest. According to the approach proposed in Foss and
Johansen (1993), Johansen and Foss (1995), the overall
model is composed by a number of local linear identified
models combined by means of a scheduling variable, the
GTload in our case, which describes their local validity
through suitably defined membership functions. The
approach considered is based on the following main steps:
• a number N (N=10) of significant values of the GTload
is selected as operating points, namely 7.5%, 15%, 25%,
40%, 50%, 57%, 60%, 65%, 75%, 100% of the full
load; this set of values is selected to cover the full range
of the variable and is more dense around 60% where the
plant exhibits a strongly nonlinear behaviour;
• the Modelica simulator is used to determine the steady
state conditions at any point;
• small step variations are imposed at each operating
point to the load, to the desuperheater flow and to the
admission valve (when it is free);
• the induced transients of the output variables of interest
are collected and subsequently used for the
identification of the transfer functions Gi(z) of linear
models with local validity. The identification procedure
is performed with the Matlab Identification toolbox.
More specifically, ARX (AutoRegressive eXogenous)
and OE (Output Error) models are used. For all the
variables, first or second order models are sufficient to
describe the transients of the outputs with satisfactory
results. An example of the results achieved is shown in
Figures 1 and 2, which show the step responses of the
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

temperature TSH and of the header stress

str due to a small identified model can reprod oduce the plant dynamics of
positive variation of the GT load at theth 75% of its value interest with satisfactory accu
ccuracy, both for local and for
and computed both with the Modelic lica model and with large variations of the working
ng point.
the identified one; Concerning the adopted proce ocedure to obtain the identified
• N membership functions are def efined and tuned, model, the number of operatinting points is important but it is
according to a trial-and-error proc rocedure, each one remarkable that the most cumb mbersome and time consuming
providing the “degree of truth” µ i ∈ [0,1], i = 1,..., N of phase concerns the definitionn and
a tuning of the membership
the i-th local model based on the current
cu value of the functions. To this regard, auto
utomatic procedures based, e.g.,
GTload. The ten membership function tions selected in this on the optimization of the quality of the open-loop
work are depicted in Figure 3. simulation results over largee transients
t could be devised to
reduce the development time.
• N models Γi , i=1,…,N, used for inter terpolation are built,
as shown in Figure 4, for any outpu tput variable. In the
figure, u is the input vector, yi is the t output of the
model, ( ui , yi ) is the input-output eq
equilibrium pair for
the i-th working point specified byy tthe corresponding
value of the GTload and µi is the valuelue of the associated
membership function.
• The output y of the overall model is ob obtained as the sum
of the outputs yi, i=1,…,N of thee local models Γi.
Therefore, the complete interpolated ted model has the
structure reported in Figure 5.
TSH
809

808.5
Figure 3: membership func
unctions (x-axis: fraction of
808
full GTload).
GTl
807.5

807
[K]

806.5

806

805.5

805

804.5
0 1000 2000 3000 4000 5000 6000 7000
time [s]

Figure 1: step responses of TSH with thehe Modelica (dotted

line) and the identified models (solid line
line) at the 75% of
GTload. Figure 4: i-th local linear
ar mmodel with membership
functtion.
7 Header Stress
x 10
9.9

9.8

9.7
[Pa]

9.6

9.5

9.4

9.3
0 1000 2000 3000 4000 5000 6000 7000
time [s]

Figure 2: step responses of the header

er stress with the
Modelica (dotted line) and the identifie
ified models (solid
line) at the 75% of GTload
load.
The transients of the temperature TSH anda of the header
stress caused by a wide ramp variation of the GTload (from
100 to the 15% of the load) computedd with
w the Modelica
model (dotted line) and with the identifie
ified one (solid line)
are reported in Figures 6 and 7. It is apparent that the Figure 5: structure of thee overall
ov model for output y.

7045
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

850
TSH
stated and where q is a vector of unknown parameters, which
have to be selected through an optimization procedure.
800
Then, the problem of computing the optimal load profile
consists of finding the value of the parameter vector q
750
together with the final time tf which solves the following
optimization problem:
[K]

700
tf

650 min J = dt
q ,t f ∫
t0
600
0 1 2 3 4 5 6 7 8 9 10
time [s] 4
x 10 subject to the constraints
Figure 6: temperature TSH computed with the Modelica
model (dotted line) and with the identified one (solid line). xɺ (t ) = f ( x(t ), d , L(t , q))
x 10
7 Header Stress
L(t f , q) ≥ LM − ε 1
11

10.5 f ( x(t f ), d , L(t f , q)) ≤ ε 2

10
g ( x(t )) ≤ 0, t 0 ≤ t ≤ t f
9.5

9
where ε1 and ε2 are arbitrary small values (ideally equal to
[Pa]

8.5
zero) and the corresponding constraints are included to
8 guarantee that at time tf the system is (almost) in stationary
7.5 conditions and has reached (almost) full load. The last
7 (vector) constraint includes, in general form, all the
6.5
constraints to be imposed to the plant variables. Specifically,
0 1 2 3 4 5 6 7 8 9 10
time [s] 4
x 10
only constraints on the stresses during the start-up precedure
have been included in this work..
Figure 7: header stress computed with the Modelica Many functions L(t,q) can be chosen, such as sigmoid
model (dotted line) and with the identified one (solid line). functions. In this work, the following Hill function has been
chosen:

4. COMPUTATION OF THE OPTIMAL LOAD PROFILE th

L(t , q) = Lm + ( LM − Lm )
With the operating conditions described in Section 2, the t + kh
h

main input that can be used to optimize the start-up is the

load of the gas turbine. The problem is then to find the where q = [h k ] is the parameter vector to be determined
GTload profile that minimizes the time required to reach full through the optimization procedure. Letting Lm=0, LM=1,
load conditions while respecting constraints on the process t0=0, tf=1, k=0.5, the Hill function for different values of h is
variables; typically, it is required that temperatures, pressures shown in Figure 8.
and stresses remain into prescribed limits for safety and
economical reasons.
1
More formally, the problem posed corresponds to a minimum
0.9
time optimal control problem and can be stated as follows.
0.8

The CCPP model can be described by the dynamic system 0.7

Hill function L(t,q)

0.6

xɺ (t ) = f ( x(t ), d , L(t )) 0.5

0.4

where x is the vector of state variables (temperatures, 0.3

pressures, …), d is the set of (constant) external inputs (feed 0.2

flow to the desuperheater, admission valve, by-pass) and L is 0.1

the GTload. 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t
Denote now by t0 the initial time instant of the start-up
procedure (conventionally it can be set t0=0) and by tf the Figure 8: Hill function with k=0.5, Lm=0, LM=1 and h=5
final one. As for the load profile, it must be selected so that (solid line), h=10 (dashed line), , h=20 (dotted line).
L(t0)=Lm and L(tf)=LM, Lm and LM being the initial and final
(full) load.
Assume now that the load is described by an increasing
function L(t,q) satisfying the boundary conditions above

7046
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

As a simple simulation model is available, the optimization 11

x 10
7 Header Stress

problem above can be solved numerically to compute the

10
unknown parameters.
9
5. VALIDATION EXPERIMENTS
8
The optimization procedure described in the previous section

[Pa]
has been implemented by imposing a number of constraints 7

on the main plant variables. In particular, the header stress 6

and the steam turbine stress have been imposed to not exceed
11e7 [Pa] and 2.25e8 [Pa], respectively. 5

The computed optimal load profile (Hill function) for a start- 4

up phase (from 15% to 100%) is reported in Figure 9. Note
that the start-up time is about 3.5 hours, but 90% of the
response is obtained in 1 hour.
This optimal load profile has been computed with the
identified interpolated local linear model and tested with the
Modelica one. The transients of some relevant variables with
this profile are reported in Figures 10 and 11. From these
results it is apparent that the response of the two models are
quite close to each other and that the computed optimal
profile on the identified model allows one to reach with the
Modelica models, the required full load conditions while
respecting the imposed state constraints. In particular, the
solution guarantees that the stress obtained during the
transient does not exceed the design value.
1
0.9
0.8
0.7
load percentage
0 2000 4000 6000 8000 10000 12000 14000
time [s]
Figure 11: header stress corresponding to the optimal
profile of GTload computed with the Modelica model
(dotted line) and with the Simulink one (solid line).
6. CONCLUSIONS
The identified model that has been designed in this project
has been used for an open loop optimization. The results of
the optimization phase allow one to consider that it is
accurate enough to be used as the prediction model of a MPC
controller, so that potential deviations of the system from its
expected value can be dealt with.
The project described in this paper can be developed in many
different ways. The first open point is the relaxation of the
operating conditions and the use also of the desuperheater
flow and the admission valve position as optimization
signals. This will require the specification of a systematic
procedure to define the operating points where to perform
local model identification and the subsequent optimization of

0.6 the corresponding membership functions, as for a three

0.5 dimension space the problem is more difficult to deal with.
0.4
Another point of interest is the selection of the function
0.3
describing the input signals. In fact, the simple choice
0.2 performed in this study may lead to very suboptimal
0.1
0 2000 4000 6000 8000 10000 12000 14000
solutions to the original problem. For example, concerning
time [s] our results (see Figure 9), while Hill functions can guarantee
fast and safe attainment of high load condition (90% of full
Figure 9: Optimal load profile.
load is attained in approximatively one hour), the final
transient phase (90% - 100% of full load) could be made
faster with a different choice of input function. In some
TSH
850 cases, for example, the cascade connection of two Hill (or
equivalent) functions could be useful to replicate the typical
800 load profile used in the plants management.
Also, additional constraints can be included in the
750
optimization problem formulation. For instance, the rate of
[K]

change of the stress level might also significantly affect the

700
plant longevity.
650 Finally, the problem considered here can be viewed as at the
higher level of a hierarchical control structure where at the
600
0 2000 4000 6000 8000 10000 12000 14000
lower level suitable controllers for the main plant variables
time [s] must be properly designed. It is necessary to study how the
uncertainty on the lower levels can be taken into account with
Figure 10: temperature TSH corresponding to the this approach at the higher level.
optimal profile of GTload computed with the Modelica
model (dotted line) and with the Simulink one (solid line). Acknowledgments: This research has been supported by the
European 7th framework STREP project ”Hierarchical and

7047
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

distributed model predictive control (HD-MPC)” contract Krüger, K., Franke, R., Rode, M.: “Optimization of boiler
number INFSO-ICT-223854. start-up using a nonlinear boiler model and hard
constraints”, Energy, Vol. 29, n. 12-15, pp. 2239-2251,
2004.
REFERENCES
Krüger, K., Rode, M., Franke, R.: “Optimal control for fast
Albanesi, C., Bossi, M., Magni, L., Paderno J., Pretolani F., boiler start-up based on a nonlinear model and
Kuehl, P., Diehl, M., (2006). Optimization of the start- considering the thermal stress on thick-walled
up procedure of a combined cycle power plant. In components.”, Proceedings of the 2001 IEEE
Proceedings of the 45th IEEE Conference on Decision International Conference on Control Applications, pp.
& Control. IEEE. 570 – 576, 5-7 Sept. 2001.
Aurora, C., (2004), Power plants: modeling, simulation and Lausterer, G.K., (1997), On-line thermal stress monitoring
control, PhD Thesis (supervisor R. Scattolini), using mathematical models. Control Engineering
University of Pavia. Practice, pages 85-90.
Aurora, C., Diehl, M., Kuehl, P., Magni, L., Scattolini, R., Leporati C, Recine M., Leporati, C., Recine M., Vaccaro A.,
(2005), Nonlinear model predictive control of combined Zangani S., (1993), “Documentazione teorica del
cycle power plants”, 16th IFAC World Congress, modulo TMSL”, ENEL internal report B1092HA0010
Prague. (in Italian).
Bausa J., Tsatsaronis G. (2001a): “Dynamic optimization of Smith, C.A., Burnham, K.J., Ham, P.A.L., Zachariah, K.J.,
startup and load-increasing processes in power plants – James, D.J.G., 1996), “Adaptive fuzzy control of
Part I: Method”, Transactions of ASME, Vol. 123, pp. combined cycle plant start up”, International
246-250. Conference on “Opportunities and Advances in
Bausa J., Tsatsaronis G., (2001b): “Dynamic optimization of International Electric Power Generation”, pp. 115 –
startup and load-increasing processes in Power Plants – 118.
Part II: Application”, Transactions of ASME, Vol. 123, Watson, W. J. (1996). The ‘success’ of the combined cycle
pp. 251-254. gas turbine. In: IEEE International Conference on
Casella, F., Leva A., (2006), “Modelling of Thermo- Opportunities and Advances in International Electric
Hydraulic Power Generation Processes using Power Generation. Durham, UK. pp. 87–92.
Modelica”, Mathematical and Computer Modelling of Xu C., Kephart, R.W., Menten, C.H., (2000), Model-based
Dynamical Systems, v. 12, n. 1, pp. 19-33. once-through boiler start-up water wall steam
Casella, F., A. Leva, (2003), “Modelica Open Library for temperature control”, Proceedings of the 2000 IEEE
Power Plant Simulation: Design and Experimental International Conference on Control Applications, pp.
Validation”, Proc. 3rd International Modelica 778 – 783.
Conference, Linköping, Sweden, pp. 41-50. URL:
http://www.modelica.org/events/Conference2003/papers
/ h08_Leva.pdf.
Casella, F., Pretolani, F., (2006), Fast Start-up of a
Combined-Cycle Power Plant: A Simulation Study with
Modelica, Modelica Conf.
Ferrari-Trecate G., Gallestey, G., Letizia, P., Spedicato, M.,
Morari, M., Antoine, M., (2002), Modelling and control
of co-generation power plants under consideration of
lifetime consumption: a hybrid system approach, 15th
IFAC World Congress, Barcelona.
Foss, B. A., Johansen, T.A: (1993). Constructing NARMAX
models using ARMAX models. Int. J. Control, Vol. 58,
pp.1125-1153.
Johansen, T.A, Foss B.A., (1995). Identification of nonlinear
system structure and parameters using regime
decomposition. Automatica, Vol. 31, pp.321-326.
Faille, D., Davelaar F., (2009), "Model based start-up
optimization of a combined cycle power plant", in
Proceedings of the IFAC Symposium on Power Plants
and Power Systems Control (IFAC PP&PSC 2009),
Tampere Hall, Finland.
Franke R., Rode M., Kruger K.: “On-line optimization of
drum boiler start-up”, Proc. of the 3rd International
Modelica Conference, 2003.
Kakimoto, N. and K. Baba (2003). Performace of gas turbine-
based plants during frequency drops. IEEE Trans. on
Power Systems 18(3), 1110–1115.

7048